The number hit my screen at 6:47 AM Jakarta time. Goldman Sachs revising wafer fab equipment (WFE) spending upward. 2027: $218 billion. 2028: $281 billion. My first reaction wasn't excitement. It was suspicion. Because in this industry, when the sell-side gets bullish, the smart money starts checking the assumptions. And these assumptions? They're built on a stack of cards that could collapse if AI capex so much as sneezes.
Let me be clear about what's actually happening here. This isn't a routine forecast adjustment. This is Goldman implicitly betting that the semiconductor industry's growth trajectory has fundamentally shifted. The implied CAGR from 2024 to 2028 lands somewhere north of 20%. For context, the last time we saw sustained WFE growth at that level was the 2016-2018 memory supercycle. And we all remember how that ended. Inventory corrections. Price crashes. Capacity oversupply that took years to digest.
But here's the thing about cycles: they don't repeat, they rhyme. And the rhyme this time is AI. Every single line of Goldman's projection flows through the AI demand channel. HBM4. 2nm GAA. CoWoS capacity. High-NA EUV. It's all AI, all the way down. The question nobody on the sell-side wants to answer: what happens if the AI capex cycle peaks in 2026, right when all this new capacity comes online?
The structural shift nobody's talking about is the packaging bottleneck.
Here's what the report gets right. The WFE spending mix is changing. Dramatically. Traditional logic foundry spending โ the TSMC-led advanced node expansion โ is no longer the sole driver. Memory is back, and it's not your father's DRAM cycle. HBM3E is already in mass production. HBM4 lands in the second half of 2025. And that requires a completely different equipment set. TSV etching. Hybrid bonding. Advanced packaging tools that didn't exist in meaningful volume five years ago.
I've been tracking this packaging angle since 2020, when I first noticed the Uniswap V2 flash loan arbitrage bots were draining liquidity pools faster than anyone could model. The parallel is structural: both are about infrastructure bottlenecks creating arbitrage opportunities. In DeFi, it was MEV bots extracting value from inefficient AMMs. In semiconductors, it's CoWoS capacity constraining AI chip supply. TSMC doubled CoWoS capacity in 2024 and still can't meet demand. That's not a demand problem. That's a supply chain that hasn't caught up to structural reality.
The equipment supply chain itself is the constraint Goldman's model glosses over.
ASML ships roughly 50-60 EUV tools per year. High-NA EUV? Single units cost over $300 million. The delivery lead time stretches 12-18 months. Now do the math on what $281 billion in WFE spending actually requires. It requires ASML to more than double EUV output. It requires Lam Research and Applied Materials to compress their 6-12 month lead times. It requires the entire precision components supply chain โ the Zeiss optics, the RF generators, the ultra-pure materials โ to scale at a pace that has never been achieved in this industry's history.
This is where my contrarian stress-testing kicks in. Based on my audit experience across multiple equipment supply chains, the bottleneck isn't demand. It never was. The bottleneck is delivery. And delivery constraints mean one of two things: either the spending gets pushed out, or it gets concentrated in the hands of whoever can secure equipment first. That's not a growth story. That's a rationing story. And rationing stories tend to end with the weakest players getting squeezed out.
The memory supercycle thesis deserves more scrutiny than it's getting.
Goldman's projection implies DRAM supply tightness persisting through 2028. That's a bold call. The last time we saw a memory supercycle โ 2017-2018 โ it was driven by smartphone demand and data center buildout. This time it's HBM. And HBM is fundamentally different. It's not a commodity product. It's a custom-engineered, high-margin product where the top three players (Samsung, SK Hynix, Micron) control essentially the entire market. That concentration is actually bullish for pricing power. But it also means the equipment spending is concentrated in fewer hands.
Here's the hidden implication: if memory becomes the dominant WFE driver, the equipment demand structure shifts. Logic fabs need EUV, deposition, and etch tools in a certain ratio. Memory fabs โ especially HBM โ need TSV etch, hybrid bonding, and advanced test equipment. The winners in this cycle won't be the same as the winners in the last cycle. Lam Research and TEL benefit disproportionately. ASML's EUV dominance becomes less relevant to the marginal dollar of WFE spend.
The China factor is the elephant in the room that Goldman's global model can't fully price.
Chinese fabs account for roughly 20-25% of global WFE spending. But here's the nuance: that spending is almost entirely in mature nodes โ 28nm and above. The equipment localization rate sits at 20-30% for mature processes and below 5% for advanced nodes. The gap isn't closing fast. EUV remains a complete blank. High-end photoresist? Still 90%+ imported from Japan. EDA tools? 95%+ from Synopsys, Cadence, and Siemens.
But here's what the bears miss: the equipment shortage is actually a tailwind for Chinese equipment makers. When global supply is tight, customers are more willing to qualify second-tier suppliers. That's how you get companies like Naura and AMEC gaining design wins in etch and deposition. It's not going to change the global WFE picture in 2026-2028. But it's a structural shift that compounds over a decade. And in this industry, a decade is one technology node cycle.
The real risk is the AI capex cliff.
Let me walk through the pre-mortem. Goldman's 2028 number assumes AI-related capex from hyperscalers โ Microsoft, Google, Meta, Amazon โ keeps growing at 40%+ annually through 2028. That's the implicit assumption. Now stress-test it. What happens if large language model commercialization disappoints? What happens if CSPs shift from buying NVIDIA GPUs to designing their own ASICs? What happens if the AI bubble โ and let's be honest, there are bubble characteristics โ deflates in 2026?
The historical precedent is instructive. In 2018, WFE spending dropped over 10% when memory prices collapsed. In 2022, it dropped again when the crypto mining boom went bust and consumer electronics demand cratered. The semiconductor industry has never sustained 20%+ WFE CAGR for four consecutive years. Not once. The capital intensity is too high, the cyclicality too severe, the herd behavior too predictable.
The contrarian play isn't shorting equipment stocks. It's understanding the timing mismatch.
Equipment stocks โ ASML, AMAT, Lam, KLA โ are trading at 25-40x forward earnings. That's the high end of their historical range. The market has already priced in the Goldman scenario. The question is whether the market has priced in the execution risk. ASML's EUV monopoly is real. But the company can't ship what it can't build. And if the AI capex cycle peaks in 2026, the 2027-2028 equipment spending gets pulled forward, creating a demand vacuum in the back half of the decade.
This is the classic semiconductor industry trap. Everyone builds for the peak. The peak arrives. Then the trough arrives. And the equipment makers โ who have the highest fixed costs and the longest lead times โ get hit hardest. The 2010s memory expansion is the template. Samsung, SK Hynix, and Micron all built aggressively. Then prices collapsed. Then they all cut capex simultaneously. The equipment makers felt the whiplash first.
The takeaway isn't to fade the equipment cycle. It's to respect the timing.
If you're positioned in equipment names, the next 12-18 months look constructive. The order books are full. The lead times are long. The pricing power is real. But the 2027-2028 numbers are where the risk lives. That's when the capacity comes online. That's when the depreciation hits. That's when the AI capex cycle โ if it follows historical patterns โ starts to roll over.
Chaos is just data we haven't decoded yet. The data says Goldman is bullish. The data also says the semiconductor industry has never sustained this growth rate. One of those data points is wrong. My money's on the historical pattern. The equipment cycle will deliver. The question is whether it delivers on Goldman's timeline or on the industry's own brutal, cyclical schedule. Watch the order books. Watch the lead times. Watch the AI capex guidance from the hyperscalers. That's where the truth lives. The rest is just sell-side noise.
Arbitrage isn't just liquidity waiting for a mirror. It's the gap between what the models predict and what the supply chain can actually deliver. And right now, that gap is wider than Goldman's spreadsheets suggest.